Hugging Face has released DFM Mimir v1, a 1-billion-parameter language model built on the Hierarchical Reasoning Model (HRM) architecture. The model is trained from scratch using only permissible post-training data, addressing the growing concern over the use of massive, often non-permissible datasets in AI development.
Mimir v1 was trained on a mixture of 161 datasets, totaling approximately 70.479 billion tokens per epoch. Despite its relatively compact size, the model outperforms the original HRM-Text 1B and competes with larger frontier models such as Qwen 3.5 4B and Gemma 4 E2B, as evaluated across 20 benchmarks covering English, Math & Code, and Danish.
Notably, Mimir v1 sets a new state of the art for Danish language performance, a significant achievement for low-resource language modeling. The model is available on the Hugging Face Hub, making it accessible for researchers and developers committed to open-source and ethically sourced data.
This release highlights a growing trend toward transparency and data ethics in AI, offering a viable path for high-performance models without compromising on data permissions.